CRL: An Efficient Autonomous Exploration Framework for Large-Scale Environments With Contrastive-Driven Reinforcement...

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Science

Background:

• Autonomous exploration in large-scale environments faces challenges in viewpoint selection due to poor feature extraction and increasing computational demands.
• Current methods often fail to address both feature extraction and computational cost issues cohesively.

Purpose of the Study:

• To develop an efficient autonomous exploration framework that overcomes limitations in viewpoint selection and computational complexity.
• To enhance the precision of optimal viewpoint selection through contrastive learning mechanisms.
• To reduce computational costs in large-scale environment exploration.

Main Methods:

• Implemented a contrastive-driven reinforcement learning framework with contrastive constraints on nodes in high-dimensional feature spaces.
• Developed specialized training rules for effective action constraints to prevent backtracking and redundant exploration.
• Introduced a novel graph rarefaction algorithm to manage computational complexity.

Main Results:

• Achieved a 6.7% shorter path length compared to state-of-the-art (SOTA) approaches.
• Demonstrated enhanced precision in optimal viewpoint selection by explicitly capturing key regional characteristics.
• Showcased robust generalization capabilities through real-world robotic experiments.

Conclusions:

• The proposed framework efficiently tackles suboptimal viewpoint selection and high computational costs in autonomous exploration.
• Contrastive learning and graph rarefaction significantly improve navigation efficiency and scalability.
• The method offers a promising solution for real-world robotic exploration in complex environments.